Computation of W directly is very expensive since it involves dot products in a very high dimensional space

The kernel function is used in the input space which equals to the dot product in the high dimentional feature space

When radial kernal function is select and input vectors have 30 features, the dimension of the new space is about 10000